On November 11, 2026, the OpenMOSS research team—led by Fudan University’s Natural Language Processing Laboratory alongside the Shanghai Innovation Institute and MOSI Intelligence—released MOSS-Live, an open-source full-duplex real-time multimodal interaction system. The system enables continuous conversational speech, concurrent listening and speaking, robust interruption handling, and decoupled background tool execution for advanced dialogue research.
Traditional spoken dialogue systems operate on a half-duplex turn-taking cadence, functioning like walkie-talkies where the user and the system must take rigid turns speaking. MOSS-Live transitions conversational AI into full-duplex operation, allowing the model to listen continuously while simultaneously synthesizing speech. This architectural shift enables spontaneous turn-taking, immediate handling of user interruptions, and natural processing of ambient audio and visual streams.
Architecture: Decoupled Foreground and Background Execution
The core technical contribution of MOSS-Live is a dual-timescale architecture that separates rapid conversational interaction from latency-heavy background computation. By separating these two processes, the system preserves low-latency spoken conversation without blocking on external tool calls or analytical reasoning.
Foreground Conversational Turn-Taking and Interruption Handling
The foreground interaction model operates on the timescale of milliseconds, governing speech synthesis, active listening, and acoustic turn-taking. It continuously evaluates audio streams to distinguish between genuine user interruptions, conversational pauses, and brief backchannel signals such as ‘mm-hmm’ or ‘uh-huh’. When a user interrupts, the foreground model yields within 1.38 seconds, adjusting its ongoing speech without losing conversational state. Furthermore, selective attention mechanisms filter out ambient background chatter and non-directed speech, ensuring the model only takes the floor when addressed.
Background Asynchronous Task Coordination
While the foreground model handles live conversation, the background model executes complex reasoning, external database queries, and API invocations. A shared coordination layer bridges the two models through persistent memory and task state. When a background task completes—such as retrieving experimental documentation or querying an academic repository—the coordination layer determines an opportune moment in the dialogue to introduce the findings, rather than abruptly interrupting the active conversational thread.
Empirical Evaluation Across Conversational Benchmarks
The OpenMOSS team evaluated MOSS-Live across six comprehensive evaluation suites spanning 36 distinct interaction metrics, comparing performance against leading proprietary and open models.
Turn-Taking Precision and Selective Attention on FDB
On the Full-Duplex Benchmark (FDB) suite, MOSS-Live demonstrated exceptional conversational timing. In FDB v1, the model recorded a 0.00% unwanted turn-over rate during backchannel listening, confirming that it does not falsely yield the floor when a user offers minor verbal acknowledgments. In user interruption scenarios, MOSS-Live achieved a post-interruption response quality score of 4.88 out of 5.00, outperforming comparable systems. In FDB v1.5 evaluations assessing selective attention, the system maintained a 98.00% disregard rate for speech directed at other individuals and a 90.00% disregard rate for ambient background sound, demonstrating robust acoustic discrimination.
Tool Execution and Speech Understanding on τ-Voice and Big-Bench Audio
Beyond acoustic turn-taking, MOSS-Live was tested on complex multi-turn task completion and speech comprehension. On the FDB v3 tool-use benchmark, the system achieved a 51.00% tool-use success rate and a 95.68% tool selection F1 score, tying for the top result among evaluated architectures. On the τ-Voice benchmark, MOSS-Live achieved a 57.93% multi-turn task success rate, combined with 98.17% backchannel selectivity and 82.16% vocal tic selectivity. Across the Big-Bench Audio evaluation, MOSS-Live reached an overall comprehension accuracy of 88.40%.
Implications for Academic Labs and Speech Dialogue Research
The release of MOSS-Live carries substantial strategic and practical implications for academic research institutions, research offices, and computer science laboratories working at the intersection of human-computer interaction and AI and machine learning research outputs.
Open Reference Models Versus Proprietary Voice Services
Prior full-duplex multimodal systems have largely remained proprietary cloud services with opaque training sets and closed inference pipelines. By releasing open weights, documentation, and evaluation scripts through their public repository, the OpenMOSS team provides university researchers with an inspectable foundation for reproducible AI experiments. Research groups investigating acoustic prosody, conversational linguistics, and multimodal alignment can analyze model behavior directly without relying on proprietary API rate limits or closed commercial terms.
Hands-Free Research Interfaces and Multimodal Lab Assistance
The decoupled background execution model provides a compelling template for academic laboratory assistance. In wet-lab or clinical research environments where researchers require hands-free interaction, an assistant powered by full-duplex architectures can listen to verbal queries, analyze video feeds from bench cameras, and dispatch data queries to electronic lab notebooks or AI research tools in the background. Because the conversational thread remains responsive while computational tasks run, researchers can engage in continuous dialogue without operational stalls. As universities establish evaluation policies on multimodal foundation models, open architectures like MOSS-Live also inform institutional oversight tracked within the Frontier AI Safety and Governance hub.
Sources
- OpenMOSS Team. (2026). MOSS-Live: A Full-Duplex Real-Time Interaction System. Hugging Face Community Article. https://huggingface.co/blog/OpenMOSS-Team/moss-live (Checked 2026-11-14).
- OpenMOSS Team. (2026). MOSS-Live: A Full-Duplex Real-Time Interaction System with Live Audio and Vision. GitHub Repository. https://github.com/OpenMOSS/MOSS-Live (Checked 2026-11-14).
- OpenMOSS Platform. (2026). MOSS-Live Realtime API Documentation and Playground. https://platform.moss.cn/en/docs/scenarios/moss-live/ (Checked 2026-11-14).
